JPH11332000A - Method for determining internal parameters of dynamic signal processing apparatus and electroacoustic apparatus - Google Patents

Abstract

(57) [Summary] To provide an electroacoustic apparatus which can freely perform sound image localization and sound image movement by setting a small number of parameters. SOLUTION: An external input signal section A for inputting a time series signal I (t) having a smooth and continuous characteristic, and a time series signal si (t) from the external input signal means A are dynamically converted. A non-linear dynamic signal processing unit B including a non-linear oscillator for processing, and a time-series signal yi (t) output from the non-linear dynamic signal processing unit B are converted into a time-series signal z (t A dynamic signal processing device is constituted by an RNN including a nonlinear signal synthesizing unit C for synthesizing the dynamic signal. Then, this dynamic signal processing device is used as an adaptive filter 22R, 22L, 24R, 24L as the sound source 2
1R, 21L, and the output of this adaptive filter
The configuration is to output from R and 11L.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a method for determining internal parameters of a dynamic signal processing device and an electroacoustic device.

[0002]

2. Description of the Related Art There are widely known methods for stimulating human sensory organs, such as stimulating the visual system with light, stimulating the auditory system with sound, and stimulating the tactile sensation with vibration. Of these stimulus methods, the presentation of warnings on moving objects such as aircraft, spacecraft, ships, railway vehicles, and automobiles is necessary to alert the pilot or driver, occupants, and passengers and deal with dangers. It is desirable to use sound presentation because it is possible to transmit an amount of information sufficient to induce an evasive action in a form that has minimal influence on the driving operation.

As an electric filter for converting a time-series signal such as voice or sound, Japanese Patent Application Laid-Open No.
One described in Japanese Patent Application No. 019085 is known.
FIG. 10 shows a FIR (Finite Impulse Response) used as a processing device for a time-series signal according to the prior art.
1 shows a filter circuit of a system.

[0004]

However, in the conventional FIR filter shown in FIG. 10, the number of taps of the convolution calculation unit has reached 4096 from [0] to [4095]. A huge amount of storage space must be secured inside the computer in order to build it on a microcomputer built into an audio device, and the size of the device is inevitable for use as an in-vehicle device. was there.

In general, an FIR as shown in FIG.
In the filter, it is necessary to calculate an impulse response in order to obtain a filter coefficient required for a sound signal filtering process such as sound image localization and speaker characteristic compensation. First, it is necessary to accurately measure the filter coefficient. For this purpose, it is necessary to measure an impulse response or a white noise response having a vibration component over the entire frequency band. However, generally available loudspeakers do not always produce a flat response over all frequency bands, and the environment in which the filter is actually used has strong reflection and absorption of sound. There is a problem that it is often difficult to design the coefficient of the FIR filter.

In addition, dynamics (control based on past sampling results) due to the structural constraint that an FIR filter generally does not have a feedback portion
There is also a problem that it is impossible to reproduce the nonlinear input-output relationship accompanied by.

[0007] The present invention has been made in view of such conventional problems, and has been described in an environment where the propagation of voice and sound is obstructed by reflection or absorption, such as inside or outside a moving object or inside or outside a building. Even below, when presented at a normal volume that is gentle on human ears and does not cause any annoyance,
It gives the listener a sense of localization of the sound image and a sense of movement of the sound image, which can be easily heard and intuitively understood, and as a result, the sound warning output which can surely draw the attention of the listener It is an object of the present invention to provide a useful electroacoustic device by utilizing.

[0008] The present invention also relates to an electroacoustic apparatus.
Generates an optimal transfer function to cancel the crosstalk that is likely to occur when performing sound image localization and sound image movement by the speaker system, and compensates for the characteristics of a real electroacoustic transducer such as a speaker, ideally. Recurrent neural network with internal dynamics (Recu Neural Network) used to construct an adaptive filter to realize the characteristics
It is an object of the present invention to provide a method for determining internal parameters of a dynamic signal processing device configured by an Rrent Neural Network (RNN).

[0009]

According to the first aspect of the present invention, there is provided an external input signal means A for inputting a time series signal I (t),
A nonlinear dynamic signal processing means B comprising a nonlinear oscillator for dynamically processing the time series signal si (t) from the external input signal means A, and a signal outputted from the nonlinear dynamic signal processing means B A dynamic signal processing device having a non-linear signal synthesizing means C for synthesizing the time-series signal yj (t) into a time-series signal z (t) having a target waveform, and comprising an RNN having internal dynamics; In order to realize a desirable input / output relationship between the time-series signal I (t) input to the dynamic signal processing device and the time-series signal z (t) output from the dynamic signal processing device, feedforward coupling coefficient with the signal input means A w ^A i and feedforward coupling coefficient possessed by the non-linear dynamic signal processing means B w ^B j
i, self coupling coefficient μi, regression coupling coefficient νi, temperature coefficient T
^B j, and constants τi time delay element Di, the nonlinear signal combining means C feedforward coupling coefficient w ^C j included in the, by a learning algorithm such as backpropagation internal parameters such as temperature coefficient T ^C of the non-linear element This allows the nonlinear input-output relationship between any time-series signals to be determined by signal processing involving dynamic dynamics.
It can be implemented using a small number of internal parameters.

According to a second aspect of the present invention, there is provided a dynamic signal processing apparatus comprising: an external input for inputting a time series signal I (t) having a smooth and continuous characteristic such as an audio signal in an audible band; Signal means A and said external input signal means A
And a time series signal yj output from the nonlinear dynamic signal processing means B for dynamically processing the time series signal si (t) from the nonlinear dynamic signal processing means B.
(t) with a non-linear signal synthesizing means C for synthesizing a time-series signal z (t) of a target waveform, and a dynamic signal processing device comprising an RNN having internal dynamics. In configuring the dynamic signal processing means B,
Measured in advance eigenmodes of vibration of the object system, its internal parameters so when they constant is a characteristic time constant of the nonlinear dynamic signal processing means, i.e., the feedforward coupling coefficient w ^B ji, self-coupling coefficient μi , The regression coupling coefficient ν
i, the temperature coefficient T ^B i, are those determined by applying a formula that is derived based on a priori knowledge and constant τi time of the delay element Di, thereby, any by a small number of nonlinear oscillators RN not only capable of realizing a nonlinear input-output relationship between time-series signals, but also responsible for internal dynamics
The coefficients of the feedforward connection, the self connection, and the regression connection of the N section can be fixed at the time of learning, and the internal parameter determination method of the dynamic signal processing device according to the third, fifth, and sixth aspects of the present invention. Learning of internal parameters to be adopted can be performed stably and promptly.

According to a third aspect of the present invention, in the method for determining an internal parameter of the dynamic signal processing device according to the first or second aspect, when the internal parameter of the nonlinear dynamic signal processing means B is determined in advance, signal input means a feedforward coupling coefficient w ^a i, are those determined by the learning algorithm 1 or more internal parameters of the feedforward coupling coefficient of the nonlinear signal combining means C w ^C j and the temperature coefficient T ^C Thereby, the external signal input means A and the nonlinear signal synthesizing means C required for realizing the non-linear input / output relationship between arbitrary time-series signals by using signal processing involving dynamics, respectively. Can be obtained through learning.

According to a fourth aspect of the present invention, in the method for determining an internal parameter of the dynamic signal processing device according to any one of the first to third aspects, the external signal input means A, the nonlinear dynamic signal processing means B and the nonlinear signal synthesizing means C After the parameters are determined,
When the time constant of the natural mode of the vibration of the target system changes,
The characteristic time constant τi of the delay element Di of the nonlinear dynamic signal processing means B is readjusted in accordance with the time constant of the eigenmode of the vibration of the target system after the change.

According to a fifth aspect of the present invention, there is provided the dynamic parameter processing method according to any one of the first to third aspects, wherein the external signal input means A, the nonlinear dynamic signal processing means B and the nonlinear signal synthesizing means C After the parameters are determined,
If the transfer function governing the vibration amplitude and the vibration waveform of the object system is changed, the external signal feedforward coupling coefficient of the input means A w ^A i, feedforward coupling coefficient w ^C i and the temperature of the nonlinear signal combining means C One of the coefficients T ^C
Alternatively, a plurality of internal parameters are readjusted in accordance with the changed transfer function.

According to a sixth aspect of the present invention, in the method for determining an internal parameter of the dynamic signal processing device according to any one of the first to third aspects, the external signal input means A, the nonlinear dynamic signal processing means B and the nonlinear signal synthesizing means C When determining the parameters and incorporating the dynamic signal processing device into the electroacoustic device, in addition to the conventional impulse response method, the desired input / output relationship of the time-series signal having the same or equivalent properties as those actually used is described. In order to realize the internal dynamics of the dynamic signal processing device, the value of the internal parameter is determined by calculation using a formula, adjustment of a time constant, or a learning algorithm.

According to the method for determining an internal parameter of a dynamic signal processing apparatus according to the present invention, the input of a time-series signal having properties equivalent to those actually used is not limited to an impulse signal or a white noise signal. By the learning based on the output relation, the internal dynamics of the dynamic signal processing device constituted by the RNN can be optimized. As a result, if the dynamic signal processing device is used as an adaptive filter of an electroacoustic device, a change in a transfer function of the sound presentation environment, such as a difference in reflection / absorption of a space for presenting an audible alarm, an individual difference in a head shape, etc. Thus, it is possible to cope with a change in characteristics of an electroacoustic transducer such as a speaker due to a change over time or a change in weather conditions.

According to a seventh aspect of the present invention, there is provided an electroacoustic apparatus comprising: an external input signal means for inputting a time series signal I (t) having smooth and continuous characteristics; A nonlinear dynamic signal processing means B comprising a nonlinear oscillator for dynamically processing the series signal si (t); and a time series signal yi (t) output from the nonlinear dynamic signal processing means B.
A dynamic signal processing device comprising an RNN provided with a non-linear signal synthesizing means C for synthesizing a sound signal with a time-series signal z (t) having a target waveform, and inputting a sound signal to the dynamic signal processing device. And an electroacoustic transducer for reproducing the output signal of the dynamic signal processing device into a sound signal and outputting the sound signal.

According to a seventh aspect of the present invention, there is provided an electroacoustic apparatus comprising: a dynamic signal processing apparatus which determines an internal parameter of each section to an optimum value using the internal parameter determining method of the dynamic signal processing apparatus according to any of the first to sixth aspects. Since it is adopted, the input signal from the sound source is converted to the optimal time-series signal reflecting the acoustic characteristics of the sound field in which the electroacoustic device is installed by the dynamic signal processing device, and output from the electroacoustic conversion device. Can be done. Therefore, by setting and adjusting the internal parameters of the dynamic signal processing device to appropriate ones, even if a sound source with a simple waveform is used, the sound, Even in the environment where sound propagation is disturbed, it is gentle to human ears and presents at normal volume without any bothersomeness. It is possible to create and output an audible alarm with a vibration waveform and vibration amplitude that can be given and heard easily, can be intuitively understood, and as a result, can surely draw the attention of the listener. .

[0018]

According to the first aspect of the present invention, since the internal parameters of the dynamic signal processing device constituted by the RNN having the internal dynamics are determined by the learning algorithm, the non-linear input between any time series signals is determined. Output relation
Signal processing with dynamic dynamics can be achieved with a small number of internal parameters.

According to the second aspect of the present invention, the internal parameters of the dynamic signal processing device constituted by the RNN having the internal dynamics are determined by applying the formula derived based on a priori knowledge. A nonlinear input / output relationship between arbitrary time-series signals can be realized by the number of nonlinear oscillators.

According to the third aspect of the present invention, when the internal parameters of the nonlinear dynamic signal processing means of the dynamic signal processing device are predetermined, the internal parameters of the external signal input means and the nonlinear signal synthesizing means are learned. The external signal input means and the nonlinear signal synthesis means required to realize the nonlinear input / output relationship between arbitrary time-series signals using signal processing with dynamic dynamics, as determined by the algorithm. Can be obtained through learning.

According to the fourth to sixth aspects of the present invention, the RNN is not limited to an impulse signal or a white noise signal, and is learned based on the input / output relationship of a time series signal having properties equivalent to those actually used. Can optimize the internal dynamics of the dynamic signal processing device composed of To be able to cope with changes in the transfer function of the sound presentation environment, such as differences in absorption and individual differences in head shape, and changes in the characteristics of electroacoustic transducers such as speakers due to changes over time and changes in weather conditions. Become.

That is, even when the electroacoustic transducer represented by the speaker does not show a flat response over the entire frequency range, or when the transfer function between the sound source and the listening point is nonlinear, By learning the input / output relationship of a time-series signal having the same or equivalent properties as those used in the above, an adaptive filter having nonlinear dynamics capable of realizing ideal characteristics can be easily constructed. As a result, even in an environment where the propagation of voice and sound is obstructed by reflection and absorption, such as inside and outside moving objects and inside and outside buildings, it is gentle to human ears and does not involve bothersome, usually Even when presented at a volume of, the listener can easily give a sense of localization of the sound image and a sense of movement of the sound image, which is easy for the listener to hear, can understand intuitively, and Can be presented with an audible alarm that can surely be sounded.

As a result, when an alarm is displayed on all moving objects such as an aircraft, a spacecraft, a ship, a railroad vehicle, and an automobile, a driving operation for calling attention of a driver or a driver, an occupant, or a passenger is performed. It is possible to transmit an audio signal having sufficient information to cause appropriate avoidance behavior in order to cope with danger while having little effect.

In particular, when performing sound image localization and sound image movement by the two-speaker system, generation of a transfer function for canceling crosstalk and compensation of the characteristics of an actual electroacoustic transducer such as a speaker, and ideal characteristics Realizing a dynamic signal processing device with internal dynamics, which is used when configuring an adaptive filter for realizing the above.

According to the seventh aspect of the present invention, the dynamic signal processing device in which the internal parameters of each section are determined to be optimal by using the internal parameter determining method of the dynamic signal processing device according to the first to sixth aspects is adapted to an adaptive filter. Since the input signal from the sound source is adopted, the dynamic signal processor converts the input signal from the sound source into an optimal time-series signal that reflects the acoustic characteristics of the sound field in which the electroacoustic device is installed. Can be output. Therefore, by setting and adjusting the internal parameters of the dynamic signal processing device to appropriate ones, even if a sound source with a simple waveform is used, the sound, Even in the environment where sound propagation is disturbed, it is gentle to human ears and presents at normal volume without any bothersomeness. It is possible to create and output an audible alarm with a vibration waveform and vibration amplitude that can be given and heard easily, can be intuitively understood, and as a result, can surely draw the attention of the listener. .

[0026]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below in detail with reference to the drawings. A method for determining internal parameters of the dynamic signal processing device according to the first embodiment of the present invention will be described. FIG.
1 shows the configuration of a dynamic signal processing device using the method for determining internal parameters of a dynamic signal processing device of the present invention. First, an embodiment of a method for determining an internal parameter of a continuous operation dynamic signal processing device described by a differential equation will be described.

The dynamic signal processing apparatus shown in FIG. 1 comprises an external input signal processing unit A for inputting a time series signal I (t) having smooth and continuous characteristics, such as an audio signal in an audible band,
A non-linear dynamic signal processing unit B comprising a non-linear oscillator for dynamically processing the time series signal si (t) from the external input signal processing unit A, and an output from the non-linear dynamic signal processing unit B Time-series signal yj (t) is converted to a time-series signal z of the target waveform.
(t), a non-linear signal synthesizing unit C for synthesizing the signal, and an RNN having internal dynamics.

The external signal input section A comprises a linear element represented by a small circle and a multiplier represented by a triangle. The external signal input section A receives an external input signal I (t) such as a voice signal or an acoustic signal, and receives a multiplier w ^A i.
Is performed, and the result is output as si (t).

[0029]

(Equation 1) The nonlinear dynamic signal processing unit B includes a linear element xi represented by a small circle.
(t), a feed-forward combination w ^B ij composed of a multiplier represented by a triangle, a self-coupled μi composed of a multiplier, a regression coupling νi composed of a multiplier, and a nonlinear element yj having an internal configuration shown in FIG.
(t) and a delay element Di having a time constant τi.

In the nonlinear dynamic signal processing section B, as shown in the following equation (2), a dynamic signal due to the action of the internal dynamics xi (t) is applied to the input signal si (t) from the external signal input section A. The nonlinear signal is subjected to nonlinear conversion and output to the nonlinear signal synthesizing unit C as a time series signal yj (t) which is also an internal representation of the RNN.

[0031]

(Equation 2) Here, the nonlinear element yj (t) can be represented by an adder Σ and a monotonically increasing saturation function generator as shown in FIG. As the f of the nonlinear element yj (t), a differentiable and monotonically increasing saturated function (for example, (1 + exp) ^-1
And tanh).

The nonlinear signal combining unit C, the feed-forward coupling comprising a multiplier expressed in triangles w ^C j, nonlinear element z, and a multiplier expressed in triangles for dynamic range adjustment of the final output z (t) alpha It is composed of The configuration of the nonlinear element z is the same as the configuration of the nonlinear element yj (t) described above and shown in FIG.

The nonlinear signal combining unit C performs the calculation of equation (3), the nonlinear dynamic signal processing unit time series signal yj received from B (t), after product-sum operation of the feedforward coupling w ^C j , A nonlinear conversion is performed by the nonlinear element z, a dynamic range of the final output is adjusted by the multiplier α, and the resultant is output to the outside of the dynamic signal processing device as the final output z (t).

[0034]

(Equation 3) In the case of a dynamic signal processing device of discrete operation described by a difference equation, the same operation equation as in the case of a continuous system is obtained.

Here, according to the first aspect of the present invention, a dynamic signal processing apparatus is constructed which changes the internal parameters of the RNN by performing learning and realizes a desired input / output relationship. This is performed as follows.

According to the nature of the time series signal I (t) input from the target system or the outside, the eigenmode of the vibration of the nonlinear oscillator to be constructed inside the nonlinear dynamic signal processing unit B is k
Pieces (in the simplest case, k = 1 can be set)
(1) First, the external signal input unit A is connected to the input time-series signal si (t).
Is formed by a feed-forward combination w ^A j composed of 2k linear elements and 2k multipliers that directly receive

(2) Next, the nonlinear dynamic signal processing unit B
The 2k and the linear element xi [2k ² + 2k] multipliers (feedforward coupling coefficient corresponds to w ^B ji is 2k ²
(The number corresponding to the self-coupling coefficient μi and the number k corresponding to the regression coupling coefficient νi), 2k nonlinear elements yj, and 2k delay elements Di. This is a finding obtained from the fact that a differential equation or a difference equation of 2k variables corresponds to k periodic vibration modes.

(3) Similarly, the nonlinear signal synthesizer C is
a feedforward combination w ^C j, 1 of k multipliers
And a single multiplier α.

A time-series signal I (t) input from outside
The characteristic time constant τi of the delay element Di included in the nonlinear dynamic signal processing section B is given in correspondence with the natural vibration mode of (1).

In the first embodiment, next, according to the first and third aspects of the present invention, a target waveform is output as an output for an input I (t) of an external time-series signal. The input / output relationship between the time-series signals is learned by the RNN so as to output the time-series signal z (t). This learning is usually included in the external signal input unit A, a feed-and-forward coupling w ^A i constituted by the multiplier is included in the nonlinear signal combining station C, a configured feedforward combined by multiplier w ^C j And the temperature coefficient T ^{C of the} non-linear element, to obtain the input / output relationship required for the dynamic signal processing device.

In FIG. 3, as an algorithm for learning these multipliers represented as hatched triangles, a dynamic signal processor is a continuous system, and its operation can be described by a differential equation. In some cases, a paper published by Pearlmutter in 1989: "Barak A.
Pearlmutter; "Learning State Space Trajectories in
Recurrent Neural Networks ", Neural Computation, V
ol.1 No.2 (Summer, 1989), 263-269 "(Reference 1) A paper published by Masaaki Sato in 1990:" Masaaki SAT
O; "A Real Time Learning Algorithm for Recurrent A
nalog Neural Networks ", Biological Cybernetics, Vo
l.62 (1990), 237-241 "(Reference 2)" Masaaki SATO; "A Learning Algorithm to Teach Spa
tiotemporal Patternsto Recurrent Neural Networks ",
Biological Cybernetics, Vol. 62 (1990), 259-263 ''
(Literature 3) A steepest descent learning algorithm using an appropriate error evaluation function described in, for example, (Document 3) is adopted.

On the other hand, the dynamic signal processing device is a discrete system,
When the operation is described by a difference equation, for example, Williams and Zipser published in 1989 as algorithms for learning a multiplier of a feedforward part in an external signal input part A and a nonlinear signal synthesis part C. Paper: “Rona
ld J. Williams and David Zipser; "A Learning Algor
ithm for Continually Running Fully Recurrent Neura
l Networks ", Neural Computation, Vol.1 No.2 (Summe
r, 1989), 270-280 ”(Reference 4) and the like, and employs a steepest descent learning algorithm using some error evaluation function.

In both the continuous system and the discrete system, even when a learning method other than the steepest descent method, for example, reinforcement learning or another learning algorithm is employed, the RN
In principle, it is possible for N to obtain the desired nonlinear input / output relationship between I (t) and z (t).

Next, a second embodiment of the present invention will be described with reference to FIG. The second embodiment corresponds to the second aspect of the present invention. Here, an embodiment of a method for determining internal parameters for a dynamic signal processing device of continuous operation described by a differential equation will be described.

In the second embodiment, first, as in the first embodiment, the nonlinear dynamic signal processor B is made to correspond to the natural vibration mode of the time-series signal I (t) input from the outside. The characteristic time constant τi of the included delay element Di is given.

Next, according to the second aspect of the present invention, the coupling coefficient w ^B ji (t) of the feedforward portion represented by the multiplier of the nonlinear dynamic signal processing unit B (in FIG. is represented as a triangle that has been subjected), self-coupling coefficient μi and regression coupling coefficient .nu.i, in addition, it is derived based on a priori knowledge of the temperature coefficient T ^B i with the output function of the non-linear element yj (t) Calculated according to the design formula.

In calculating these quantities, for example, the present inventors filed a patent application in 1993 and published in 1995: “Recurrent Neural Network (Japanese Patent Laid-Open No. 7-084974)” (Reference 5). A preliminary paper at an international conference published by the present inventors in 1995: "Kazumi YOSHIDA and Norimasa KISHI;" On the A
rchitechture of Recurrent Neural Network as a Chao
s 1 / f Generator ", PROCEEDINGS of the International
Joint Conference of the Fourth IEEE International
Conference on Fuzzy Systems and the Second Intern
ational Fuzzy Engineering Symposium (FUZZ-IEEE / IFE
S '95), (YOKOHAMA, March 24, 1995), 1963-1970 ''
(Reference 6) or a US patent filed by the present inventors in the United States in 1994 and patented in 1997: "APPARATUS BASED ON N-VARIABLE"
UNLIMITED RECURRENT ADJUSTABLE NETWORK (USP5
671, 336) "(Reference 7).

Next, in the second embodiment, according to the third aspect of the present invention, a time series signal input I
The input / output relationship between the time-series signals is acquired by learning so that the time-series signal z (t) having the target waveform is output as the output of the RNN for (t). This learning is usually included in the external signal input unit A, a feed-and-forward coupling w ^A i constituted by a multiplier, a nonlinear signal combining unit C
Included in the subjects feed against forward coupling w ^C i and a multiplier for determining a dynamic range α formed by the multiplier, to optimize the input-output relationship of the dynamic signal processing apparatus.

In the algorithm for learning the multiplier represented by the hatched triangle in FIG. 3, the dynamic signal processing device is a continuous system, and its operation can be described by a differential equation. For example, those described in Literatures 1 to 3 exemplified in the description of the first embodiment are employed. When the dynamic signal processing device is a discrete system and its operation is described by a difference equation, for example, the one described in Reference 4 exemplified in the description of the first embodiment is employed. I do.

When a learning method other than the steepest descent method, for example, reinforcement learning or another learning algorithm is employed to calculate the internal parameters of the external signal input section A and the nonlinear signal synthesizing section C, the desired nonlinear input method is also obtained. It is possible to obtain an output relationship.

Next, a third embodiment of the present invention will be described with reference to FIG. The third embodiment corresponds to the fourth aspect of the present invention. According to the third embodiment, in the dynamic signal processing device shown in FIG. 4, the delay element Di represented by a hatched rectangle in the nonlinear dynamic signal processing unit B is changed over time or changes in weather conditions. It is characterized in that the adjustment is performed when the natural vibration mode of the target system changes due to the above.

That is, the time-series signal I input from the outside
The characteristic time constant τi of the delay element Di included in the nonlinear dynamic signal processing unit B is adjusted according to the change in the natural vibration mode of (t). At the time of this adjustment, the final output z (t) of the dynamic signal processing device at that time is followed in accordance with the second embodiment.
An error function, such as a square error, between the error and the target output d (t) is obtained, and the coupling coefficient w to be moved in the case of the normal error propagation method
^This is performed by changing the time constant τi instead of Bji (t).

At this time, the coupling coefficients w ^B ji (t) of the feedforward portion, the coupling coefficients μi of the self-coupling portion, which are included in the nonlinear dynamic signal processing section B and are each represented by a multiplier, temperature coefficient T ^B j with the output function f of the coupling coefficient νi and nonlinear elements yi regression binding moiety to be kept fixed. The multiplier α for adjusting the dynamic range in the nonlinear signal synthesizing unit C is determined by learning or by a formula.

Next, a fourth embodiment of the present invention will be described with reference to FIG. The fourth embodiment corresponds to the fifth aspect of the present invention. In the fourth embodiment, in the dynamic signal processing device shown in FIG. 5, the multiplier w in the external signal input section A represented by a hatched triangle is used.
^A i, the multiplier w ^C j in the nonlinear signal synthesizing unit C, the multiplier α, and the temperature coefficient T ^C of the nonlinear element included in the nonlinear signal synthesizing unit C are changed over time, individual differences, changes in weather conditions, etc. Thus, additional learning is performed when the transfer function of the target system changes.

That is, the time series signal I input from the outside
The multiplier w ^A i in the external signal input unit A, the multiplier w ^C j, α in the nonlinear signal synthesis unit ^C , and the nonlinear signal synthesis unit performing additional learning temperature coefficient T ^C of the non-linear element z contained C. At the time of this additional learning, following the above-described second embodiment,
The square error between the final output z (t) of the dynamic signal processing device at that time and the target output d (t) is obtained, and these internal parameters are changed by a steepest descent algorithm such as an error back propagation method. .

Also in this case, the coupling coefficient w ^B ji (t) of the feedforward portion, the coupling coefficient μi of the self-coupling portion, and the regression coupling portion represented by the multiplier in the nonlinear dynamic signal processing section B are also used. And the output function f of the nonlinear element yi
Specific time constant τi temperature coefficient T ^B j and the delay element Di due to the leave fixed.

Next, a fifth embodiment of the present invention will be described with reference to FIG. The fifth embodiment corresponds to the sixth aspect of the present invention. The fifth embodiment shows a device for adaptively identifying an inverse function of a nonlinear transfer function of a target system when a dynamic signal processing device is used as an adaptive filter F of an electroacoustic device.

When compensating the transfer function of the electroacoustic transducer 11 such as a speaker, the sound source signal delayed through the delay circuit 15 and the time series signal picked up by the measuring microphone 13 in the anechoic chamber 12 are used. , An error signal e (t) is generated from the difference, and the internal parameters of the adaptive filter F configured by the dynamic signal processing device are changed according to a design formula or a learning algorithm. As the nonlinear transfer function of the target system at that time, the transfer function of the electroacoustic transducer 11: HSP The space transfer function in the anechoic chamber 12: H0 The transfer function combining the microphone 13 for measurement and the microphone preamplifier (not shown) : HMIC Transfer function with integrated amplifier (not shown): H
AMP, etc., but generally it can be considered that the HSP represents the transfer function of the whole system,
Inside the adaptive filter F, the inverse dynamics of HSP: HSP ^-1 may be obtained.

Next, a sixth embodiment of the present invention will be described with reference to FIGS. The sixth embodiment also corresponds to the invention of claim 6. 7 and 8 show the inverse of the nonlinear transfer function HSP of the electroacoustic transducer 11 of the target system according to the fifth embodiment shown in FIG. 6 with respect to the adaptive filter F constituted by the dynamic signal processor. Function HSP ^-1
Is already obtained, an apparatus for obtaining a head-related transfer function by adaptive identification is further shown.

Referring to FIGS. 7 and 8, compensation of a generally nonlinear head-related transfer function when a sound emitted from an electroacoustic transducer 11 such as a speaker propagates through a space including the pinna. think of. In this case, cabin 1
4, an error signal e (t) is generated from the difference between the time-series signal picked up by the measurement microphone 13 installed in the ear canal and the time-series signal of the sound source in the space of the actual presentation environment such as the inside. Parameters of the adaptive filter F configured by the dynamic signal processing device are obtained by a design formula or a learning algorithm.

At this time, as the nonlinear transfer function of the target system in FIG. 7, the transfer function of the electroacoustic transducer 11: the head-related transfer function in the HSP presentation environment: HLR, HLL, HR
L, HRR Transfer function combining microphone for measurement 13 and microphone preamplifier (not shown): Transfer function combining HMIC integrated amplifier (not shown): H
The HSP has already been compensated for by the inverse function HSP- ¹ obtained by the fifth embodiment shown in FIG.

Therefore, it can be generally considered here that the four head-related transfer functions HLL, HLR, HRL, and HRR are representative of the transfer function of the entire system. Thus, in the same manner as in the fifth embodiment, the internal parameters are obtained for each transmission path by using a design formula or a learning algorithm, so that the adaptive filter F of the dynamic signal processing device including the RNN is obtained. Inside, forward dynamics of the head-related transfer function: HLR, HRL Inverse dynamics of the head-related transfer function: HLL ⁻¹ , HRR ⁻¹ are acquired.

That is, by using these four nonlinear transfer functions, it is possible to design an adaptive filter F having internal parameters for compensating for nonlinear dynamics derived from individual differences and the acoustic transfer environment.

Next, an electroacoustic apparatus according to a seventh embodiment of the present invention will be described with reference to FIG. The seventh embodiment corresponds to the seventh aspect of the present invention. The electro-acoustic apparatus shown in FIG. 9 is for realizing a three-dimensional sound image presentation, and uses an adaptive filter F designed using the apparatus of the fifth and sixth embodiments described above with sound sources of the two speakers 11R and 11L. It is configured to be inserted on the signal side.

That is, the left and right sound sources R and L are input via the audio signal amplifying devices 21R and 21L, respectively, and the adaptive filters 22R and 22L for crosstalk adjustment constituted by the dynamic signal processing device and the balance controller B _R , B _L ,
Adaptive filters 23R and 23 for compensating speaker characteristics
In this configuration, signals are input to the left and right speakers 11R and 11L installed in the passenger compartment 14 through the L.

Internal Dynamics of Adaptive Filter 22R:
A _R = −H _RL · H _RR ⁻¹ Internal dynamics of adaptive filter 22L: A _L = −H _LR ·
H _LL ^−1, which is a combination of the internal dynamics obtained in the fifth and sixth embodiments.

According to the electroacoustic apparatus having such a configuration,
Practical voice / sound presentation can be performed, in which the sound image is clear and the effects of sound image localization and sound image movement are easy to grasp.

[Brief description of the drawings]

FIG. 1 is a block diagram of a dynamic signal processing device used in a first embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a nonlinear element in the dynamic signal processing device.

FIG. 3 is a block diagram of a dynamic signal processing device used in a second embodiment of the present invention.

FIG. 4 is a block diagram of a dynamic signal processing device used in a third embodiment of the present invention.

FIG. 5 is a block diagram of a dynamic signal processing device used in a fourth embodiment of the present invention.

FIG. 6 is a block diagram of a device for optimizing internal parameters of a dynamic signal processing device used in a fifth embodiment of the present invention in order to use it as an adaptive filter.

FIG. 7 is a block diagram of a first apparatus for optimizing internal parameters of a dynamic signal processing apparatus used in a sixth embodiment of the present invention so as to use the apparatus as an adaptive filter.

FIG. 8 is a block diagram of a second apparatus for optimizing an internal parameter of a dynamic signal processing apparatus used in a sixth embodiment of the present invention so as to be used as an adaptive filter.

FIG. 9 is a block diagram of an electroacoustic apparatus according to a seventh embodiment of the present invention.

FIG. 10 is an explanatory diagram of a conventional example.

[Explanation of symbols]

 A external signal input section B nonlinear dynamic signal processing section C nonlinear signal synthesis section F adaptive filter

Claims (7)

[Claims] 1. An external input signal means for inputting a time series signal, and a non-linear dynamic signal processing means comprising a non-linear oscillator for dynamically processing a time series signal from the external input signal means. A dynamic signal constituted by a recurrent neural network (RNN) having a nonlinear signal synthesizing means for synthesizing a time series signal output from the nonlinear dynamic signal processing means into a time series signal having a target waveform. The external signal input means for realizing a desirable input / output relationship between a time series signal input to the dynamic signal processing apparatus and a time series signal output from the dynamic signal processing apparatus. Feed-forward coupling coefficient, feed-forward coupling coefficient, self-coupling coefficient, regression coupling coefficient and temperature coefficient of the nonlinear oscillator of the nonlinear dynamic signal processing means, Dynamic signal processing, wherein one or more internal parameters of a time constant of a delay element included in the dynamic signal processing means, a feedforward coupling coefficient of the nonlinear signal synthesizing means, and a temperature coefficient are determined by a learning algorithm. A method for determining internal parameters of the device. 2. An external input signal means for inputting a time-series signal, and a nonlinear dynamic signal processing means comprising a nonlinear oscillator for dynamically processing a time-series signal from the external input signal means. A dynamic signal processing device comprising an RNN having a nonlinear signal synthesizing means for synthesizing the time series signal output from the nonlinear dynamic signal processing means into a time series signal having a target waveform. In determining the internal parameters of the non-linear dynamic signal processing means, the eigenmodes of the vibration of the target system are measured in advance, and their feeds are set so that their time constants become the characteristic time constants of the non-linear dynamic signal processing means. One or more internal parameters of the forward coupling coefficient, the self coupling coefficient, the regression coupling coefficient, the temperature coefficient, and the time constant of the delay element of the nonlinear dynamic signal processing means are determined based on a priori knowledge. Internal parameter determining method of dynamic signal processing apparatus characterized by determining by applying the formulas derived Te. 3. When the internal parameters of the non-linear dynamic signal processing means are determined in advance, a feed-forward coupling coefficient of the external signal input means, a feed-forward coupling coefficient of the non-linear signal synthesizing means, and a temperature coefficient. 3. The method according to claim 1, wherein one or more internal parameters are determined by a learning algorithm. 4. When the time constant of the natural mode of vibration of the target system changes after the internal parameters of the external signal input means, the nonlinear dynamic signal processing means and the nonlinear signal synthesizing means are determined, the nonlinear dynamic The time constant of the delay element of the dynamic signal processing means is readjusted in accordance with the time constant of the eigenmode of the vibration of the target system after the change.
The method for determining internal parameters of a dynamic signal processing device according to any one of the above. 5. When a transfer function governing a vibration amplitude or a vibration waveform of a target system changes after internal parameters of the external signal input means, the nonlinear dynamic signal processing means, and the nonlinear signal synthesis means are determined, One or more internal parameters of the feedforward coupling coefficient of the external signal input means, the feedforward coupling coefficient of the nonlinear signal synthesizing means, and the temperature coefficient are readjusted in accordance with the changed transfer function. The method for determining an internal parameter of a dynamic signal processing device according to claim 1. 6. A method for determining internal parameters of the external signal input means, the nonlinear dynamic signal processing means and the nonlinear signal synthesizing means, and incorporating the dynamic signal processing apparatus into an electroacoustic apparatus in addition to an impulse response method. In order to realize the desired input / output relationship of the time-series signal having the same or equivalent properties as those used in the dynamic signal processing device as internal dynamics, calculation by a formula, adjustment of a time constant, or learning algorithm 4. The method for determining an internal parameter of a dynamic signal processing device according to claim 1, wherein the value of the parameter is determined. 7. An external input signal means for inputting a time series signal, and a non-linear dynamic signal processing means comprising a non-linear oscillator for dynamically processing a time series signal from said external input signal means. A dynamic signal constituted by a recurrent neural network (RNN) comprising: a time series signal output from the nonlinear dynamic signal processing means; and a nonlinear signal synthesizing means for synthesizing a time series signal having a target waveform. An electroacoustic apparatus comprising: a processing device; a sound source that inputs a sound signal to the dynamic signal processing device; and an electroacoustic conversion device that reproduces an output signal of the dynamic signal processing device into a sound signal and outputs the sound signal.